Skip to content

[https://nvbugs/5721907][fix] AutoDeploy: improve numerical stability of flashinfer attention test #10467

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lucaslie merged 1 commit into from
Jan 7, 2026
Merged

[https://nvbugs/5721907][fix] AutoDeploy: improve numerical stability of flashinfer attention test #10467

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 0 additions & 1 deletion tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -379,7 +379,6 @@ accuracy/test_llm_api_pytorch.py::TestNemotronUltra::test_fp8_prequantized[tp8-c
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[cutlass-auto] SKIP (https://nvbugs/5596343)
unittest/_torch/speculative/test_spec_gate.py::test_spec_gate_e2e SKIP (https://nvbugs/5710045)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=TRTLLM-mtp_nextn=0-ep4-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] SKIP (https://nvbugs/5721661)
unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/test_flashinfer_attention_op.py::test_flashinfer_attention_op_context_input_pos[cuda-dtype0-4-8-seq6] SKIP (https://nvbugs/5721907)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] SKIP (https://nvbugs/5722629)
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_2gpus[cutlass-two_model-overlap_scheduler] SKIP (https://nvbugs/5702826)
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[cutlass-two_model-overlap_scheduler] SKIP (https://nvbugs/5702826)
Expand Down
11 changes: 11 additions & 0 deletions tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/test_flashinfer_attention_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -574,6 +574,17 @@ def test_flashinfer_attention_op_context_input_pos(seq, batch_size, n_heads, dty
(MAX_BATCH_SIZE, MAX_SEQ_LEN, N_HEADS, D_HEAD), dtype=DTYPE, device=device
)

# Initialize the prefilled portion of the cache with random data
# This simulates a chunked prefill scenario where previous chunks have already
# populated the cache at positions 0:PREFILL_SEQ_LEN
if PREFILL_SEQ_LEN > 0:
k_cache[0:BATCH_SIZE, 0:PREFILL_SEQ_LEN, :, :] = torch.randn(
BATCH_SIZE, PREFILL_SEQ_LEN, N_HEADS, D_HEAD, dtype=DTYPE, device=device
)
v_cache[0:BATCH_SIZE, 0:PREFILL_SEQ_LEN, :, :] = torch.randn(
BATCH_SIZE, PREFILL_SEQ_LEN, N_HEADS, D_HEAD, dtype=DTYPE, device=device
)

# make sure planner is initialized
workspace = torch.empty(128 * 1024 * 1024, dtype=torch.uint8, device=device)
_GlobalFlashInferPlanner.init_workspace(workspace)
Expand Down
Loading